Your First AI application

Going forward, AI algorithms will be incorporated into more and more everyday applications. For example, you might want to include an image classifier in a smart phone app. To do this, you'd use a deep learning model trained on hundreds of thousands of images as part of the overall application architecture. A large part of software development in the future will be using these types of models as common parts of applications.

In this project, you'll train an image classifier to recognize different species of flowers. You can imagine using something like this in a phone app that tells you the name of the flower your camera is looking at. In practice you'd train this classifier, then export it for use in your application. We'll be using this dataset from Oxford of 102 flower categories, you can see a few examples below.

The project is broken down into multiple steps:

We'll lead you through each part which you'll implement in Python.

When you've completed this project, you'll have an application that can be trained on any set of labeled images. Here your network will be learning about flowers and end up as a command line application. But, what you do with your new skills depends on your imagination and effort in building a dataset. For example, imagine an app where you take a picture of a car, it tells you what the make and model is, then looks up information about it. Go build your own dataset and make something new.

Import Resources

Load the Dataset

Here you'll use tensorflow_datasets to load the Oxford Flowers 102 dataset. This dataset has 3 splits: 'train', 'test', and 'validation'. You'll also need to make sure the training data is normalized and resized to 224x224 pixels as required by the pre-trained networks.

The validation and testing sets are used to measure the model's performance on data it hasn't seen yet, but you'll still need to normalize and resize the images to the appropriate size.

Explore the Dataset

Label Mapping

You'll also need to load in a mapping from label to category name. You can find this in the file label_map.json. It's a JSON object which you can read in with the json module. This will give you a dictionary mapping the integer coded labels to the actual names of the flowers.

There are 2 ways to find the flower name: eithrr using class_names[str(label+1)] or using get_label_name(label).

From the above dictionary which is from an external json file, we can see that 73 corresponds to water lily. Below is a figure of water lily with label 72! I have check the naming convention, for some other images too, and noticed that, 1 should be added to the label number class_names[str(label+1)].

However, if we use dataset_info.features['label'], we get the correct label.

Create Pipeline

Data Augmentation

The ImageDataGenerator class has three methods: .flow is for When your data comes in the form of a NumPy array. .flow_from_directory is for When your data lives locally on your machine with a separate folder containing images of each class. .flow_from_dataframe is for When your data lives locally on your machine in a single folder source. So here, flow() should be used.

flow() requires seperate X and y, so, the follwoing function is used to get X and y from data.

Note: I have not used the augmented data to train the model!

Validation and test data generators without augmentation and shuffling:

Build and Train the Classifier

Now, it's time to build and train the classifier.

Use MobileNet pre-trained model:

The MobileNet pre-trained model is used from TensorFlow Hub to get the image features. Build and train a new feed-forward classifier using those features.

Things are done:

Use Cov2D:

Predict with untrained model:

Plot along the probabilities:

Train the model from MobileNet:

Train the CovNet model:

See how many are predicted correctly:

Plot the loss and accuracy values achieved during training for the training and validation set:

Testing your Network

It's good practice to test your trained network on test data, images the network has never seen either in training or validation. This will give you a good estimate for the model's performance on completely new images. You should be able to reach around 70% accuracy on the test set if the model has been trained well.

To test the network, I use model.evaluate, and not model.predict. Because:

.predict() generates output predictions based on the input you pass it.

.evaluate() computes the loss based on the input you pass it, along with any other metrics that you requested in the metrics param when you compiled your model.

Compare model that used MobileNet and the other that is built using CovNet:

Save the Model

Now that your network is trained, save the model so you can load it later for making inference. In the cell below save your model as a Keras model (i.e. save it as an HDF5 file).

Model can be saved as bellow, howver, the best modle is already saved abovem, as I have used callbacks=[save_best].

Load the Keras Model

Load the Keras model you saved above.

The current model is the bset_model.h5 that is saved when model was fitted to the data because I have used .fit and callbacks=[save_best]. So, the three models including bset_model.h5, time.h5, the reloded model saved_keras_model_filepath are the same.

To reload a model that is from Transfer Learning technigue from hub, custom_objects={'KerasLayer':hub.KerasLayer} should be used.

Compare current model with the loaded model:

Inference for Classification

Now you'll write a function that uses your trained network for inference. Write a function called predict that takes an image, a model, and then returns the top $K$ most likely class labels along with the probabilities. The function call should look like:

probs, classes = predict(image_path, model, top_k)

If top_k=5 the output of the predict function should be something like this:

probs, classes = predict(image_path, model, 5)
print(probs)
print(classes)
> [ 0.01558163  0.01541934  0.01452626  0.01443549  0.01407339]
> ['70', '3', '45', '62', '55']

Your predict function should use PIL to load the image from the given image_path. You can use the Image.open function to load the images. The Image.open() function returns an Image object. You can convert this Image object to a NumPy array by using the np.asarray() function.

The predict function will also need to handle pre-processing the input image such that it can be used by your model. We recommend you write a separate function called process_image that performs the pre-processing. You can then call the process_image function from the predict function.

Image Pre-processing

The process_image function should take in an image (in the form of a NumPy array) and return an image in the form of a NumPy array with shape (224, 224, 3).

First, you should convert your image into a TensorFlow Tensor and then resize it to the appropriate size using tf.image.resize.

Second, the pixel values of the input images are typically encoded as integers in the range 0-255, but the model expects the pixel values to be floats in the range 0-1. Therefore, you'll also need to normalize the pixel values.

Finally, convert your image back to a NumPy array using the .numpy() method.

To check your process_image function we have provided 4 images in the ./test_images/ folder:

The code below loads one of the above images using PIL and plots the original image alongside the image produced by your process_image function. If your process_image function works, the plotted image should be the correct size.

Once you can get images in the correct format, it's time to write the predict function for making inference with your model.

Inference

Remember, the predict function should take an image, a model, and then returns the top $K$ most likely class labels along with the probabilities. The function call should look like:

probs, classes = predict(image_path, model, top_k)

If top_k=5 the output of the predict function should be something like this:

probs, classes = predict(image_path, model, 5)
print(probs)
print(classes)
> [ 0.01558163  0.01541934  0.01452626  0.01443549  0.01407339]
> ['70', '3', '45', '62', '55']

Your predict function should use PIL to load the image from the given image_path. You can use the Image.open function to load the images. The Image.open() function returns an Image object. You can convert this Image object to a NumPy array by using the np.asarray() function.

Note: The image returned by the process_image function is a NumPy array with shape (224, 224, 3) but the model expects the input images to be of shape (1, 224, 224, 3). This extra dimension represents the batch size. We suggest you use the np.expand_dims() function to add the extra dimension.

Sanity Check

It's always good to check the predictions made by your model to make sure they are correct. To check your predictions we have provided 4 images in the ./test_images/ folder:

In the cell below use matplotlib to plot the input image alongside the probabilities for the top 5 classes predicted by your model. Plot the probabilities as a bar graph. The plot should look like this:

You can convert from the class integer labels to actual flower names using class_names.

First plot the image with all probabilities for all classess:

Find the top_k classes for the above example: